Info-YOLO: A Novel Multiscale Feature Enhancement Architecture for Remote Sensing Object Detection

The detection of dense, small objects in remote sensing imagery is significantly challenged by high altitude, complex backgrounds and ultra-high resolution, which often leads to false positives and false negatives, notably in scenes with dense and small objects. To address the aforementioned challenges, this paper proposes Info-YOLO, a novel algorithm designed to reliably identify small and densely packed targets against complex backgrounds. Our initial step is to propose an Efficient Channel Attention mechanism and apply it to C2f and SPPF in the backbone network, called Feature Enhancement and Extraction Module (FEEM) and ECA-enhanced Spatial Pyramid Pooling Fast (ECSPPF). FEEM enhances the multiscale feature extraction capability, and ECSPPF alleviates information loss associated with multistep pooling. In addition, to alleviate the problem of inaccurate detection caused by overlapping objects, we employ an improved Bidirectional Feature Pyramid Network (BiFPN) for its superior feature fusion ability, replacing the conventional Path Aggregation Network (PANet) and achieving more effective integration of multiscale features with superior performance. Furthermore, to further boost the detection accuracy for small targets, a Swin Transformer block is inserted at the transition point linking the network's neck and the prediction head. Our model achieves a new state-of-the-art mAP of 95.3% on the same dataset, surpassing all contemporary methods. To facilitate reproducibility and further research, the source code is publicly available at: https://github.com/linyuesummer/Info-YOLO-paper-code.